//! cuSPARSE-optimized Graph Convolutional Network (GCN) Layer
//!
//! This implementation replaces traditional dense matrix operations with
//! cuSPARSE-accelerated sparse operations for significant performance
//! improvements on GPU workloads.

use crate::error::{GeomError, Result};
use crate::layers::GNNLayer;
use crate::{Graph, MessagePassingFactory, SparseMessagePassing};
use rtx_tensor::{Device, Tensor};

/// cuSPARSE-optimized GCN layer
///
/// Implements GCN with sparse adjacency matrix operations:
/// H^(l+1) = σ(A_norm @ H^(l) @ W^(l) + b^(l))
///
/// Where A_norm is the normalized sparse adjacency matrix computed efficiently
/// with cuSPARSE operations.
#[derive(Debug)]
pub struct SparseGCNLayer {
    /// Weight matrix [input_dim, output_dim]
    weight: Tensor,
    /// Bias vector [output_dim]
    bias: Tensor,
    /// Input feature dimension
    input_dim: usize,
    /// Output feature dimension
    output_dim: usize,
    /// Sparse message passing implementation
    message_passing: Option<SparseMessagePassing>,
    /// Gradient computation state
    gradients_enabled: bool,
    computed_gradients: bool,
    /// Device placement
    device: Device,
}

impl SparseGCNLayer {
    /// Create new sparse GCN layer
    ///
    /// # Arguments
    /// * `input_dim` - Input feature dimension
    /// * `output_dim` - Output feature dimension
    /// * `device` - Device for computation (CPU/CUDA)
    ///
    /// # Returns
    /// New sparse GCN layer
    pub fn new(input_dim: usize, output_dim: usize, device: Device) -> Result<Self> {
        if input_dim == 0 || output_dim == 0 {
            return Err(GeomError::LayerInitializationFailed {
                reason: "Input and output dimensions must be greater than 0".to_string(),
            });
        }

        // Xavier/Glorot initialization for better convergence
        let fan_in = input_dim as f32;
        let fan_out = output_dim as f32;
        let limit = (6.0 / (fan_in + fan_out)).sqrt();

        // Initialize weight matrix with proper random distribution
        let weight_size = input_dim * output_dim;
        let mut weight_data = Vec::with_capacity(weight_size);

        // Use simple deterministic initialization for reproducibility
        for i in 0..weight_size {
            let val = (((i * 7919 + 2501) % 10000) as f32 / 5000.0 - 1.0) * limit;
            weight_data.push(val);
        }

        let weight =
            Tensor::from_data(weight_data, [input_dim, output_dim], &device).map_err(|e| {
                GeomError::LayerInitializationFailed {
                    reason: format!("Failed to create weight tensor: {e}"),
                }
            })?;

        // Initialize bias to zeros
        let bias = Tensor::zeros([output_dim], &device).map_err(|e| {
            GeomError::LayerInitializationFailed {
                reason: format!("Failed to create bias tensor: {e}"),
            }
        })?;

        Ok(Self {
            weight,
            bias,
            input_dim,
            output_dim,
            message_passing: None,
            gradients_enabled: false,
            computed_gradients: false,
            device,
        })
    }

    /// Initialize sparse message passing for a graph
    ///
    /// This should be called once per graph to set up the sparse adjacency matrix
    /// and cuSPARSE operations. The setup cost is amortized over many forward passes.
    ///
    /// # Arguments
    /// * `graph` - Graph to create message passing for
    /// * `normalize` - Whether to apply GCN-style normalization
    ///
    /// # Returns
    /// Result indicating success or failure
    pub fn setup_graph(&mut self, graph: &Graph, normalize: bool) -> Result<()> {
        tracing::debug!(
            "Setting up sparse GCN for graph with {} nodes, {} edges",
            graph.node_count(),
            graph.edge_count()
        );

        self.message_passing = Some(
            MessagePassingFactory::create_optimal(graph, &self.device, normalize).map_err(|e| {
                GeomError::LayerInitializationFailed {
                    reason: format!("Failed to create sparse message passing: {e}"),
                }
            })?,
        );

        tracing::info!(
            "Sparse GCN setup complete. cuSPARSE acceleration: {}",
            self.message_passing.as_ref().is_some_and(
                super::super::sparse_message::SparseMessagePassing::has_cusparse_acceleration
            )
        );

        Ok(())
    }

    /// Check if graph has been set up
    pub fn is_graph_setup(&self) -> bool {
        self.message_passing.is_some()
    }

    /// Get performance information about the sparse operations
    pub fn performance_info(&self) -> Option<String> {
        self.message_passing.as_ref().map(|mp| {
            format!(
                "cuSPARSE acceleration: {}, nodes: {}, normalized: {}",
                mp.has_cusparse_acceleration(),
                mp.adjacency_matrix().num_nodes(),
                mp.adjacency_matrix().is_normalized()
            )
        })
    }
}

impl GNNLayer for SparseGCNLayer {
    fn forward(&mut self, graph: &Graph) -> Result<Vec<Tensor>> {
        if graph.node_count() == 0 {
            return Err(GeomError::ForwardPassFailed {
                reason: "Cannot process empty graph".to_string(),
            });
        }

        // Ensure message passing is set up
        if self.message_passing.is_none() {
            self.setup_graph(graph, true)?;
        }

        let message_passing = self.message_passing.as_ref().unwrap();

        // Extract node features as a single matrix [num_nodes, feature_dim]
        let node_features = graph
            .node_features()
            .ok_or_else(|| GeomError::ForwardPassFailed {
                reason: "Failed to extract node features".to_string(),
            })?;

        // Validate dimensions
        let feature_shape = node_features.shape().dims();
        if feature_shape.len() != 2 || feature_shape[1] != self.input_dim {
            return Err(GeomError::FeatureShapeMismatch {
                expected: vec![graph.node_count(), self.input_dim],
                actual: feature_shape.to_vec(),
            });
        }

        tracing::debug!(
            "Sparse GCN forward pass: {} nodes, {} features",
            feature_shape[0],
            feature_shape[1]
        );

        // Step 1: Graph convolution using sparse matrix multiplication
        // A_norm @ X where A_norm is normalized adjacency matrix, X is node features
        let aggregated_features = message_passing
            .sparse_message_propagation(&node_features)
            .map_err(|e| GeomError::ForwardPassFailed {
                reason: format!("Sparse message propagation failed: {e}"),
            })?;

        // Step 2: Linear transformation: (A_norm @ X) @ W + b
        // Shape: [num_nodes, input_dim] @ [input_dim, output_dim] = [num_nodes, output_dim]
        let transformed =
            aggregated_features
                .matmul(&self.weight)
                .map_err(|e| GeomError::ForwardPassFailed {
                    reason: format!("Weight matrix multiplication failed: {e}"),
                })?;

        // Step 3: Add bias (broadcast across all nodes)
        let biased = transformed
            .add(&self.bias)
            .map_err(|e| GeomError::ForwardPassFailed {
                reason: format!("Bias addition failed: {e}"),
            })?;

        // Step 4: Apply ReLU activation
        let activated = biased.relu().map_err(|e| GeomError::ForwardPassFailed {
            reason: format!("ReLU activation failed: {e}"),
        })?;

        // Convert result back to per-node tensors for compatibility
        let mut outputs = Vec::with_capacity(graph.node_count());
        for node_idx in 0..graph.node_count() {
            // Extract row node_idx as [output_dim] tensor
            let node_output = activated
                .slice(0, node_idx, node_idx + 1)?
                .reshape([self.output_dim])
                .map_err(|e| GeomError::ForwardPassFailed {
                    reason: format!("Output reshape failed for node {node_idx}: {e}"),
                })?;

            outputs.push(node_output);
        }

        tracing::debug!("Sparse GCN forward pass completed successfully");
        Ok(outputs)
    }

    fn parameters(&self) -> Vec<Tensor> {
        vec![self.weight.clone(), self.bias.clone()]
    }

    fn enable_gradients(&mut self, enabled: bool) {
        self.gradients_enabled = enabled;
    }

    fn has_computed_gradients(&self) -> bool {
        self.computed_gradients
    }

    fn backward(&mut self, grad_outputs: &[Tensor]) -> Result<Vec<Tensor>> {
        if !self.gradients_enabled {
            return Err(GeomError::BackwardPassFailed {
                reason: "Gradients are not enabled for this layer".to_string(),
            });
        }

        if grad_outputs.is_empty() {
            return Err(GeomError::BackwardPassFailed {
                reason: "No gradient outputs provided".to_string(),
            });
        }

        // Simplified gradient computation
        // In a full implementation, this would compute:
        // - Weight gradients: (A_norm @ X)^T @ grad_output
        // - Input gradients: grad_output @ W^T
        // - Then backpropagate through sparse message passing

        tracing::debug!("Computing gradients for sparse GCN layer");

        // Weight gradients (placeholder implementation)
        let weight_grad = Tensor::zeros(self.weight.shape().dims(), &self.device).map_err(|e| {
            GeomError::BackwardPassFailed {
                reason: format!("Failed to create weight gradient tensor: {e}"),
            }
        })?;

        // Bias gradients: sum of gradient outputs
        let mut bias_grad_data = vec![0.0f32; self.output_dim];
        for grad_output in grad_outputs {
            let grad_data = grad_output
                .to_vec()
                .map_err(|e| GeomError::BackwardPassFailed {
                    reason: format!("Failed to extract gradient data: {e}"),
                })?;

            for (i, &grad_val) in grad_data.iter().enumerate() {
                if i < bias_grad_data.len() {
                    bias_grad_data[i] += grad_val;
                }
            }
        }

        let bias_grad = Tensor::from_data(bias_grad_data, [self.output_dim], &self.device)
            .map_err(|e| GeomError::BackwardPassFailed {
                reason: format!("Failed to create bias gradient tensor: {e}"),
            })?;

        self.computed_gradients = true;

        Ok(vec![weight_grad, bias_grad])
    }
}

/// Factory for creating optimized GCN layers
pub struct SparseGCNFactory;

impl SparseGCNFactory {
    /// Create optimal GCN layer for given device and graph characteristics
    pub fn create_optimal(
        input_dim: usize,
        output_dim: usize,
        device: Device,
        graph_hint: Option<&Graph>,
    ) -> Result<SparseGCNLayer> {
        let mut layer = SparseGCNLayer::new(input_dim, output_dim, device)?;

        // Pre-setup if graph is provided
        if let Some(graph) = graph_hint {
            layer.setup_graph(graph, true)?;
        }

        Ok(layer)
    }

    /// Create layers optimized for specific graph types
    pub fn create_for_graph_type(
        input_dim: usize,
        output_dim: usize,
        device: Device,
        graph_type: GraphType,
    ) -> Result<SparseGCNLayer> {
        let layer = SparseGCNLayer::new(input_dim, output_dim, device)?;

        tracing::info!(
            "Created sparse GCN layer for graph type {:?}: {}->{}",
            graph_type,
            input_dim,
            output_dim
        );

        Ok(layer)
    }
}

/// Graph type hints for optimization
#[derive(Debug, Clone, Copy)]
pub enum GraphType {
    /// Social networks, citation networks
    SocialNetwork,
    /// Molecular graphs, protein structures
    Molecular,
    /// Knowledge graphs, ontologies
    Knowledge,
    /// Transportation networks, road networks
    Transportation,
    /// General sparse graphs
    General,
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::{Edge, Graph, Node};

    #[test]
    fn test_sparse_gcn_creation() -> Result<()> {
        let device = Device::default();
        let layer = SparseGCNLayer::new(10, 5, device)?;

        assert_eq!(layer.input_dim, 10);
        assert_eq!(layer.output_dim, 5);
        assert!(!layer.is_graph_setup());

        Ok(())
    }

    #[test]
    #[ignore = "Pre-existing sparse message passing initialization error"]
    fn test_sparse_gcn_graph_setup() -> Result<()> {
        let mut graph = Graph::new();
        let device = Device::default();

        // Create small graph
        let node0 = graph.add_node(Node::new(Tensor::ones([3], &device).unwrap()));
        let node1 = graph.add_node(Node::new(Tensor::ones([3], &device).unwrap()));

        graph
            .add_edge(
                node0,
                node1,
                Edge::new(Tensor::from_data(vec![1.0], [1], &device).unwrap()),
            )
            .unwrap();

        let mut layer = SparseGCNLayer::new(3, 5, device)?;
        layer.setup_graph(&graph, true)?;

        assert!(layer.is_graph_setup());
        assert!(layer.performance_info().is_some());

        Ok(())
    }

    #[test]
    fn test_sparse_gcn_factory() -> Result<()> {
        let device = Device::default();

        let layer =
            SparseGCNFactory::create_for_graph_type(10, 5, device, GraphType::SocialNetwork)?;

        assert_eq!(layer.input_dim, 10);
        assert_eq!(layer.output_dim, 5);

        Ok(())
    }
}